Hacker Runs Massive AI on a Gaming PC, No Expensive Hardware Needed
This could save you thousands on hardware and keep your data private.
A developer built an inference engine called Xolotl, based off Strata, that runs DeepSeek-V4.1-Flash — all 552B of it — on one gaming PC: an RTX 4090 with 24GB VRAM, a Ryzen 7 5700X, 90GB of DDR4, and an NVMe drive. The weights are spread across VRAM, pinned RAM, and NVMe, and the engine pulls in the experts each token actually needs. The developer reports 27 tok/s decode, 180–320 tok/s prefill, 512K context, 21/21 HumanEval, 10/10 tool calls, and 3/3 agentic bug-fix loops, all through the real server with thinking on. It also runs GLM-5.3-Flash, Qwen3.8-Flash-Next, and Qwen3.8-27B from the same engine.
The GitHub is https://github.com/Indras-Mirror/Xolotl and the current quant — uncensored V4.1, 152 GiB — is at https://huggingface.co/IndrasMirror/DeepSeek-V4.1-Flash-GSQ-Uncensored. The developer is downloading the new DeepSeek-V4.1-Flash-1017, which is 510GB and slow to download on their internet. Once it is down, they plan to abliterate it and quantise it for Xolotl. They say it has the same architecture and config as V4.1, with identical config and tokenizer, so it should run at the same throughput, but the weights are all different, so it is a full re-run of calibration and quant rather than a quick patch. They hope to have everything up in a couple of days.
- A new tool called Xolotl runs a massive 552-billion-parameter AI model on a single gaming PC.
- It requires an RTX 4090 graphics card and 90GB of RAM, but no expensive server hardware.
- An uncensored version is coming in a few days, and the tool is free on GitHub.
Why It Matters
This could make powerful AI accessible to more people, saving money and increasing privacy.